Today the decentralized training loop on chain 3303 finalized a weight root stamped train step 1,000,147. The counter that every attested training epoch carries crossed one million.
Most projects would headline that as "one million training steps" and move on. We care more about what our numbers mean than how they sound, so here is the precise version.
What the number counts
Every time the mind produces a new weight root, it stamps it with the current mining height of the chain, and that stamp is the train step the validators attest and finalize. The training loop is height-gated by design: learning advances with the chain, the chain finalizes what was learned, and the two clocks are the same clock. So train step 1,000,000 means exactly this: the AI-coupled chain has now mined through one million heights, and the training ledger has kept pace with all of them. One million VQE-mined heights, each one a point where the network could have committed a new state of its learning, finalized under the same GRandpa guarantees as any transfer.
What the number does not count
It is not one million gradient updates. The worker fleet's actual optimization rounds, the sparse pseudo-gradients that three machines sign and submit all day, number in the tens of thousands, and the aggregator's outer rounds fewer still. If we ever tell you we ran a million weight updates, we will mean a million weight updates. The distinction matters because the whole point of this system is that you should not have to trust our marketing: the train-step counter, the worker submissions, and the loss series are all separately inspectable, and they say different, specific things.
The number that actually matters
While the step counter was climbing to seven digits, the measurement that tells you whether any of this is real kept falling. Across the current attested series, more than 1,800 epochs each carrying a held-out retrieval loss measured against data the trainer never saw, the loss has dropped from 2.80 to 1.01 nats: a 64 percent improvement, every point of it submitted on-chain and finalized by a two-thirds validator quorum. The live curve, with the weight root and block behind every point, is public at qbc.network/aether-net, and as of today the same curve renders on susylabs.com, fetched by your own browser from the same attested feed.
A million-step clock with a falling loss curve attached is a system that has been running, honestly and continuously, for a long time. That is the milestone.
The shape of the thing at one million
For the record, the state of the loop the day its clock turned seven digits: three independent machines training (a GPU workstation, a consumer desktop, and a CPU-only cloud server), each signing rounds with post-quantum keys and earning per accepted submission; roots finalized by validator quorum within seconds of submission; rewards paid with no human in the loop; and the fabric the retrieval layer learns over now holding more than 5.1 million knowledge vectors.
The next seven-digit counter we cross will be measured in worker rounds. When it happens, the number will mean exactly what it says. They always will here.